Citation

Li Y, Ko Y, Lee W. Fire Safety J. 2022; 132: e103629.

Copyright

(Copyright © 2022, Elsevier Publishing)

DOI

10.1016/j.firesaf.2022.103629

PMID

unavailable

Abstract

This paper proposes a novel hybrid model for flashover prediction in a compartment fire based on visual information from RGB images that are the same as those captured by regular vision cameras. The proposed model was developed as a research tool to study the feasibility of predicting flashover based on RGB vision data. This model consists of sub-modules with data-based methods using Deep Neural Networks and knowledge-based methods using fire safety science and mathematical model. One of the crucial features of the proposed model is enabled by a novel Dual-Attention Generative Adversarial Network that is developed in this study for the vision-to-infrared conversion process. The model and the overall procedure were validated against published test data from a compartment fire.

RESULTS show that the proposed model achieved promising performance, which also shows the potential to monitor the constant changes in a room fire through continuous processing images of flame and smoke.

Language: en

Keywords

Deep neural networks; Dual-attention generative adversarial network; Fire safety science; Flashover; Image processing